Tilion vs. Browserbase (2026)
Tilion returns a typed block signal on every fetch, so a research agent reads `status: "blocked"` instead of a CAPTCHA page; Browserbase leaves block detection to your own code and gates its strongest stealth mode behind its top self-serve tier.
Introduction and methodology note
Tilion and Browserbase both sell hosted browser sessions that agents attach to over the Chrome DevTools Protocol with Playwright or Puppeteer. Both can render JavaScript, take screenshots, and drive pages at scale. They diverge on the question this page is about: what your agent receives when a site fights back. Tilion's fetch, scrape, and extract routes render the page in a real stealth browser and return a typed outcome, including an explicit blocked status with the detected wall vendor and the recovery it attempted. Browserbase's Fetch API does not execute JavaScript, caps responses at 5 MB with a 60-second timeout, and documents no retry or escalation behavior, per its own docs.
Browserbase is the incumbent: 50 million sessions, Stagehand, Director, a Session Inspector debugger, and a $40M Series B. Tilion is newer, Y Combinator-backed, and currently ranks first on both independent ComputeSDK boards and the 22 October 2026 Browser Arena run. This page compares them on the block-signal problem specifically, then covers the rest of the surface.
Methodology note: Tilion figures on this page come from Tilion's published comparison documentation, product pages, and repositories, linked inline where a public page exists (see tilion.dev for docs and contact). Browserbase figures are stated as plain text from Browserbase's own published docs and pricing pages, which we do not hyperlink; where a figure comes from Tilion's comparison documentation rather than a page we loaded, we say so. All accuracy, latency, and protected-site figures are self-reported by the vendor that ran them, and last verified 23 October 2026. Test both products on your own target sites before deciding.
TL;DR
- Tilion's fetch routes return a typed outcome (
status,content_ok,detected,self_healed,recovery) with an exit escalation ladder from datacenter through residential and mobile IPs, so a block is a machine-readable signal, never a challenge page in your context window. - Browserbase is stronger on the surrounding ecosystem: Stagehand, Director, the Session Inspector debugger, four regions to Tilion's three, Selenium support, and SOC 2 Type II with a HIPAA BAA on its top self-serve tier.
- On protected-site success, Tilion's own 71-site harness scored Tilion 86% versus Browserbase 42%; the independent Fortress engine test loaded 86.4% of 92 protected sites versus Camoufox at 74.5%.
Quick-comparison table
| Dimension | Tilion | Browserbase | |---|---|---| | Fetch with real browser rendering | Yes: POST /v1/fetch, /v1/scrape, /v1/extract render in a live stealth session | Fetch API has no JavaScript execution, 5 MB max, 60-second timeout (per Browserbase docs) | | Typed block signal | status: ok, blocked, empty, no_results, redirected_off_target, incomplete; plus detected wall vendor and mechanism | Not published; no documented block-detection response field | | Automatic retry on block | Exit escalation ladder: datacenter, ISP, residential, mobile | Not published; no documented retry or escalation behavior | | Stealth mode | Always on; stealth: false returns 400 stealth_always_on | Stealth Mode listed as No on Free, Basic on Developer and Startup, Advanced on Scale; advancedStealth deprecated in favor of verified: true (Scale only, beta) | | Session length cap | ttl up to 604,800 seconds (7 days) | 6 hours on self-serve tiers, 15 minutes on Free (per Browserbase docs) | | Crawl endpoint | POST /v1/crawl, up to 100 pages and 5 hops | No crawl endpoint | | Search endpoint | POST /v1/search, real SERPs across DuckDuckGo, Bing, Google, Yahoo, up to 100 results with snippets | Search wraps Parallel and Exa, 25 results max, no snippet or page-content field | | CAPTCHA solving | On demand via POST /v1/solve (hCaptcha, reCAPTCHA v2/v3, Turnstile, Cloudflare interstitials) | Automatic, 5 to 30 seconds, types not enumerated in docs | | Regions | us, de, sg | Four regions | | ComputeSDK lifecycle, 23 Oct 2026 | 1st, 98.4 | 4th, 95.0 | | ComputeSDK throughput, 23 Oct 2026 | 1st, 90.9 | 2nd, 81.5 | | Browser Arena total median, 22 Oct 2026 | 156 ms, 99/100 runs | 594 ms, 100/100 runs | | Pricing | Not published; Tilion Cloud is in beta with a waitlist | Published self-serve tiers; Stealth Mode gated by tier |
Granular differentiator table
The differentiator that matters for a research agent is what comes back when a protected site pushes back. This table compares the block-handling surface.
| Capability | Tilion | Browserbase | |---|---|---| | Typed fetch outcome | status (ok, blocked, empty, no_results, redirected_off_target, incomplete), content_ok, detected (wall vendor family, category, mechanism), self_healed, recovery | Not published; Fetch API returns content with no documented block classification | | Challenge-page filtering | Fetch routes render in a real browser and classify the result before returning it, so a challenge never reaches the model as page text | Not published; agent code must detect blocks itself | | IP escalation on block | Automatic ladder: datacenter, ISP, residential, mobile | Not published | | Live IP rotation | rotate-ip over REST on a running session | Not published as a single-call REST action | | Network recon | POST /v1/inspect returns first-party JSON and XHR endpoints behind a page | No equivalent endpoint | | Deterministic replay | POST /v1/task derives selectors once and replays with zero model tokens; POST /v1/workflow replays from a recording | Rerunnable scripts start a model and use tokens, per Browser Use docs; Browserbase's Stagehand cache works on Browserbase browsers only and is tied to a Stagehand process | | Session observability after exit | GET /v1/session/{id}/status returns a terminal record for an hour after the session ends | Session Inspector with video replay, CDP event logs, network and console tabs |
Browserbase is cheaper to adopt if you already run Stagehand, and its Session Inspector is the better debugging surface in this category. On the block-signal question itself, Browserbase documents no equivalent of the typed outcome or the escalation ladder.
Why teams choose Tilion over Browserbase
#### The model never reads a challenge page
Tilion's fetch, scrape, and extract routes load the URL in a live stealth session and classify the result server-side before returning it. The response carries status (one of ok, blocked, empty, no_results, redirected_off_target, incomplete), content_ok, detected (the wall vendor family, category, and mechanism), and self_healed with the recovery attempted. A research agent that calls POST /v1/fetch gets either the article or a structured reason it did not get the article. There is no path where Cloudflare interstitial text enters the context window as if it were the source.
#### Blocks self-heal instead of failing the run
When a fetch comes back blocked, Tilion escalates the exit through a ladder: datacenter, then ISP, then residential, then mobile. The recovery field reports what happened, and self_healed tells you the retry succeeded without your orchestration code knowing an anti-bot system exists. Browserbase's docs describe no retry or escalation behavior; per its docs, its Fetch API does not execute JavaScript at all, so the first line of defense is missing.
#### Independent benchmarks back the session layer
Per ComputeSDK's published results, its independent weekly run of 23 October 2026 ranked Tilion first on lifecycle (98.4 versus Browserbase 95.0) and first on throughput (90.9 versus 81.5). Per Browser Arena's published 22 October 2026 run, Tilion's total median across create, connect, navigate, and release was 156 ms against Browserbase's 594 ms, though Browserbase completed 100 of 100 runs where Tilion completed 99. On protected-site success, Tilion's own 71-site harness (self-run, so read it with that in mind) scored Tilion 86% and Browserbase 42%; that figure is published on Tilion's blog and is a vendor claim. Browserbase's own benchmark material covers Stagehand speed, not session-level block rates, and it publishes no protected-site success figure.
If your agent pipeline is losing hours to challenge pages, the fastest way to evaluate the difference is to join the Tilion Cloud waitlist at tilion.com and run the same fetch list through both.
Best use cases for Tilion
#### Research agents reading protected pages
The core loop is POST /v1/fetch or POST /v1/extract with a schema, so the agent receives either structured content or a typed blocked signal with the detected vendor. The escalation ladder and rotate-ip handle the retry. This is the exact failure mode the prompt-free version of this page is about: an agent that summarizes a block page as the article.
How Browserbase compares: Browserbase can do the reading through a full session and Stagehand, but its Fetch API does not execute JavaScript and its docs define no block classification, so block detection is your code's job.
#### High-volume crawling without a model in the loop
POST /v1/crawl handles up to 100 pages with 5 hops, and POST /v1/task derives selectors on the first call and replays them with zero model tokens, repairing on page change. For a 10,000-page crawl, that is the difference between paying for tokens on every page and paying once.
How Browserbase compares: Browserbase has no crawl endpoint; you compose one from sessions and Stagehand, and Stagehand's cache is tied to a Stagehand process.
#### Long-lived sessions and persistent logins
Tilion sessions run with ttl up to 604,800 seconds, seven days, with cookie import via POST /v1/auth/import and a credential vault with an approval_required policy. Logged-in research across a week does not require re-authentication choreography.
How Browserbase compares: Browserbase caps sessions at 6 hours on self-serve tiers and 15 minutes on Free, per its own docs.
#### Deterministic agent tasks without model spend
POST /v1/agent exposes 22 deterministic tasks (login, fill_form, paginate_collect, extract_table, download_file, multi_step_checkout, dismiss_popups, and more) that run with no model tokens. POST /v1/plan gives a pre-run estimate before you commit.
How Browserbase compares: Browserbase Agents run a natural-language task on a hosted model, which is more flexible and costs tokens per run.
Best use cases for Browserbase
#### Teams already invested in Stagehand and Director
Browserbase wrote Stagehand (25.8k GitHub stars, MIT licensed, TypeScript, Python, and Go) and Director, its no-code builder. If your codebase is Stagehand-native, the SDK swap to Tilion is real work, and Browserbase's ecosystem is a genuine reason to stay.
#### Debugging-heavy agent development
The Session Inspector, with video replay, CDP event logs, network and console tabs, and a Stagehand tab, is the best debugging surface in this category, and Browserbase offers four regions to Tilion's three, Selenium support, hosted Agents, x402 pay-per-use, and SOC 2 Type II with a HIPAA BAA on its top self-serve tier. Compliance-constrained teams should weigh that seriously.
Platform overview of Tilion
#### Stealth-hardened sessions
Every Tilion Cloud session is stealth-hardened; stealth cannot be switched off, and stealth: false returns 400 stealth_always_on. Sessions offer residential egress, region and fingerprint-persona configuration, and attach over CDP. See tilion.com for the product overview and the current beta waitlist.
#### Typed fetch, scrape, and extract
POST /v1/fetch, /v1/scrape, and /v1/extract render the page in a real browser and return the typed outcome described above. A minimal call:
``python import os, httpx base = os.environ["TILION_API_URL"] # your Tilion API endpoint r = httpx.post(f"{base}/v1/fetch", headers={"Authorization": f"Bearer {key}"}, json={"url": target_url}) out = r.json() if out["status"] == "blocked": print(out["detected"], out["recovery"]) else: print(out["content"]) ``
#### Crawl and search
POST /v1/crawl walks up to 100 pages with 5 hops. POST /v1/search runs real SERPs across DuckDuckGo, Bing, Google, and Yahoo, returning up to 100 results with snippets and rich blocks. Both return the same typed outcome fields, so a blocked crawl page is a signal, not a surprise.
#### Deterministic tasks and workflows
POST /v1/agent runs 22 deterministic tasks with no model tokens. POST /v1/task derives selectors on first use and replays them, repairing on page change. POST /v1/workflow replays a flow from a recording. POST /v1/plan estimates a task before it runs.
#### Fortress, the open-source engine
Tilion's Fortress engine is a Chromium fork that fixes fingerprints in the browser's C++ layer: 81+ patches covering GPU, screen, fonts, and timezone, with headless and headed modes presenting the same fingerprint. On Fortress's published 92-site test (28 September 2026, headless, datacenter IPs, no proxies, methodology and harness in the repo), the engine loaded 86.4% of protected sites versus Camoufox at 74.5% and stock Chrome at 34.2%. v3 is free for developers with no caps; production keys for larger organizations are described at fortress.tilion.com/pricing. Install is pip install tilion-fortress or docker run --rm -p 9222:9222 tilion/fortress:latest (Docker Hub).
#### MCP server for coding agents
Fortress ships a local MCP server with 29 tools, free, no account required, letting Claude, Codex, Cursor, Windsurf, and Cline read blocked pages, extract fields, and drive forms. Setup instructions are in the repo's AGENTS.md.
How to choose decision matrix
| If you need | Choose | |---|---| | A machine-readable block signal on every fetch, so the model never sees a challenge page | Tilion | | Full JavaScript rendering on fetch with automatic IP escalation | Tilion | | Crawling and SERP search as first-class endpoints | Tilion | | Sessions longer than 6 hours, up to 7 days | Tilion | | Stagehand, Director, and an existing Browserbase codebase | Browserbase | | The best debugging and replay UI (Session Inspector) | Browserbase | | Four regions, Selenium, x402 pay-per-use | Browserbase | | SOC 2 Type II and a HIPAA BAA on a self-serve tier | Browserbase | | An open-source, self-hosted stealth engine you can run locally | Tilion's Fortress engine, which is free for developers | | Lowest session-start latency on current independent benchmarks | Tilion (156 ms versus 594 ms total median, Browser Arena, 22 October 2026) |
Migration guide
#### Switching from Browserbase to Tilion
1. Create a session. Replace client.browsers.create() with POST {base}/v1/session, where base is your Tilion API endpoint. Use an Idempotency-Key header for safe retries. 2. Attach over CDP. The Browserbase cdpUrl becomes Tilion's connect_url. There is also a native playwright_url endpoint.
Before (Browserbase):
``python b = client.browsers.create() p.chromium.connect_over_cdp(b.cdp_url) ``
After (Tilion):
``python s = httpx.post(f"{base}/v1/session", headers={"Authorization": f"Bearer {key}"}, json={"ttl": 86400}).json() p.chromium.connect_over_cdp(s["connect_url"]) ``
3. Map the settings. browserSettings.context maps to mode: "persistent" plus saved logins. proxies: true maps to residential: true. geolocation maps to geo. For protected pages, call POST /v1/fetch first and attach to the page it opened. 4. Watch for friction. Browserbase's Session Inspector has no Tilion equivalent at parity; Tilion offers a JPEG screencast over WebSocket, WebRTC video, and noVNC for full_browser: true sessions, which covers replay but not the tabbed inspector. Browserbase supports four regions to Tilion's three (us, de, sg). Selenium has no Tilion equivalent; use Playwright or Puppeteer over CDP. 5. Decide on a partial migration. If Stagehand is central to your product, keep Browserbase for the agent layer and route fetches through Tilion. That is a supported end state, not a failure.
FAQ
Does Tilion ever return a challenge page as content?
No. The fetch, scrape, and extract routes render the page in a live stealth session and classify the result before returning it. A blocked fetch returns status: "blocked" with detected naming the wall vendor family, category, and mechanism, plus self_healed and recovery describing the escalation attempt.
What happens when a site blocks a Tilion fetch?
Tilion escalates the exit through datacenter, ISP, residential, and mobile IPs. If the retry succeeds, self_healed is true and the content comes back in the same response flow. If it does not, you get the typed blocked status and can decide what to do.
Is Tilion faster than Browserbase?
On ComputeSDK's independent run of 23 October 2026, Tilion scored 98.4 versus Browserbase's 95.0 on lifecycle and 90.9 versus 81.5 on throughput. On Browser Arena's 22 October 2026 run, Tilion's total median was 156 ms versus 594 ms, with Browserbase completing 100 of 100 runs and Tilion 99. Both benchmarks are self-reported by their own harnesses; test on your own workload.
Which is better at getting past anti-bot walls?
Tilion's own 71-site harness scored Tilion 86% and Browserbase 42%, but that harness is self-run, so treat it as a vendor claim. Browser Use's own Stealth Bench V2 (25 September 2026, 100 sites) scored Browserbase 78.5% in standard mode and did not include Tilion, and the two tests use different site lists. Run both against your own targets.
What does Browserbase do better?
Stagehand and its SDK ecosystem, Director, the Session Inspector debugger, four regions to Tilion's three, Selenium support, hosted Agents, x402 pay-per-use, and SOC 2 Type II with a HIPAA BAA on its top self-serve tier. Browserbase also completed 100 of 100 runs on the 22 October Browser Arena run where Tilion completed 99.
How much does Tilion Cloud cost?
Pricing is not published. Tilion Cloud is in beta with a waitlist at tilion.com. The open-source Fortress engine is separate: v3 is free for developers with no caps, and production keys for larger organizations are described at fortress.tilion.com/pricing.
If your research agent keeps summarizing block pages, join the Tilion Cloud waitlist at tilion.com and point the same fetch list at a typed-outcome API instead.